Spectral Hashing With Semantically Consistent Graph for Image Indexing

The ability of fast similarity search in a large-scale dataset is of great importance to many multimedia applications. Semantic hashing is a promising way to accelerate similarity search, which designs compact binary codes for a large number of images so that semantically similar images are mapped to close codes. Retrieving similar neighbors is then simply accomplished by retrieving images that have codes within a small Hamming distance of the code of the query. Among various hashing approaches, spectral hashing (SH) has shown promising performance by learning the binary codes with a spectral graph partitioning method. However, the Euclidean distance is usually used to construct the graph Laplacian in SH, which may not reflect the inherent distribution of the data. Therefore, in this paper, we propose a method to directly optimize the graph Laplacian. The learned graph, which can better represent similarity between samples, is then applied to SH for effective binary code learning. Meanwhile, our approach, unlike metric learning, can automatically determine the scale factor during the optimization. Extensive experiments are conducted on publicly available datasets and the comparison results demonstrate the effectiveness of our approach.